Investigating landslide data balancing for susceptibility mapping using generative and machine learning models

With the development and application of machine learning, significant advances have been made in landslide susceptibility mapping. However, due to challenges in actual field landslide investigations, current landslide susceptibility mapping is usually characterized by insufficient landslide samples (positive samples) and low reliability of non-landslide samples (negative samples). Considering Lianghe County in Yunnan Province, China, as an example, this paper aims to research the effectiveness of three oversampling models in generating positive samples for landslides: Conditional Tabular Generative Adversarial Networks (CTGAN), Generative Adversarial Networks (GAN), and the traditional Synthetic Minority Oversampling Technique (SMOTE) algorithms. Additionally, three machine learning methods, including 1D Convolutional Neural Network-Long Short-Term Memory Neural Network (CNN-LSTM), Random Forest (RF), and Gradient Boosting Decision Tree (GBDT) classifiers, are used for landslide susceptibility assessment. We also devise a non-landslide data (negative samples) screening method utilizing a self-trained support vector machine within a semi-supervised framework. The results show that by training on the dataset after negative sample screening, the AUC values for the 1D-CNN-LSTM, RF, and GBDT models have shown significant improvement, increasing from (0.778, 0.869, 0.849) to (0.837, 0.936, 0.877). Compared with the original training set, the prediction accuracy of the three machine learning models is improved after training on the augmented data by CTGAN, GAN, and SMOTE models. The RF model, augmented with 200 positive samples generated by CTGAN, achieves the highest prediction accuracy in the study (AUC = 0.962). The 1D CNN-LSTM model achieves its highest prediction accuracy (AUC = 0.953) when augmented with 200 positive samples from GAN. Similarly, the GBDT model reaches its highest prediction accuracy (AUC = 0.928) when augmented with 200 positive samples created by SMOTE. In addition, the spatial distribution of data indicates that the data generated by the generative adversarial model exhibits higher diversity, which can be used for landslide susceptibility assessment.